Show HN on August 2 was all agent cleanup, not agent capability.

Four projects in 24 hours: CCN ("nuking the crap Claude left in the codebase"), Galda (return to Claude Code/Codex sessions without digging through git), evidence-to-skill (a gate between untrusted sources and agent skills), and claude-copy. None has traction individually: most at 2-7 points. The clustering is the finding. This week's unsolved problem is auditing and recovering from agent output, not producing more of it.